ABSTRACT: Accurate segmentation of solid tumors is challenging, due to their heterogeneous shape, extent, and location, as well as their appearance variation caused by the diversity of medical imaging. Manual annotation is tedious, prone to misinterpretation, human error, and observer bias. All these factors hinder further image analysis towards understanding tumor radio-phenotypes, predicting clinical outcomes, and monitoring progression patterns. Computational competitions have been seeking optimal advanced computational segmentation algorithms (ACSAs) for specific abnormalities, by pooling multi-institutional data together and benchmarking ACSAs from international groups. Along these lines, we have been successfully leading the organization of the International Brain Tumor Segmentation (BraTS) challenge, since 2012, towards a publicly-available pooled dataset of 542 multi-parametric MRI scans of glioma patients from 19 institutions. In the summarized analysis of all BraTS results, we have shown that although individual ACSAs do not outperform the gold standard agreement across expert clinicians, their fusion does outperform it, in terms of both accuracy and consistency across subjects. Towards the wider application of these ACSAs, in 2017 we created the BraTS algorithmic repository to make available Docker containers of individual ACSAs, created by BraTS participants. However, fusion of these ACSAs is still out of reach for clinical researchers, as there is no graphical user interface (GUI) to facilitate it, and execution of such algorithms requires substantial computational background by the user. Furthermore, although competitions such as BraTS have shown promise, they cannot easily scale due to the requirement of pooling patient data from multiple institutions to a single location, that often faces legal, privacy, and data-ownership concerns. These concerns motivate distributed learning solutions, where the data are always retained within their institutions. We have been investigating such solutions to avoid the current paradigm of multi-institutional collaboration, i.e., data-sharing, and considering their potential multi-institutional adoption, with respect to privacy, scalability, and performance, we found federated learning (FL) to be most appropriate. In FL, each institution trains a model and shares it (without patient data) with an aggregation server, which then integrates institutional models in parallel and distributes back a consensus model. In this proposal, we focus on developing the open- source Federated Tumor Segmentation (FeTS) platform, which with a user-friendly GUI will aim at i) bringing pre- trained models of various ACSAs and their fusion closer to clinical experts, and ii) allowing secure multi- institutional collaborations via FL to improve these pre-trained models without sharing patient data, thereby overcoming legal, privacy, and data-ownership challenges. Successful completion of this project will lead to an easy-to-use potentially-translatable tool enabling easy, fast, objective, repeatable and accurate tumor segmentation, without requiring a computational background by the user, and while facilitating further analysis of tumor radio-phenotypes towards accelerating discovery.

NARRATIVE: Successful completion of this project will lead to a clinically translatable, easy-to-use software tool that offers a) pre-trained tumor segmentation models and their fusion, to perform better than experts, and b) a federated learning framework facilitating secure multi-institutional collaborations to improve these pre-trained models without the need to share patient data, thereby overcoming legal, privacy, and data-ownership challenges, towards accelerating research of cancer radio-phenotypes.